Sooner or later, every business that automates with AI reaches the same decision. The tool can now do more than answer questions. It can reply to customers, update records, schedule payments and send messages. How much of that should happen without anyone looking?

Most owners pick one of two extremes. Some approve nothing, because the point of automation is to stop checking things. Others approve everything, because that feels safe. Both fail, and the second one fails in a way that is easy to miss.

This guide gives you a way to decide human in the loop AI questions one action at a time, instead of choosing a single policy for the whole workflow.

Why “approve everything” isn’t the safe option with human in the loop AI

Requiring a person to approve every action sounds responsible. In practice, a person who sees a hundred approval requests a day stops reading them. The approval becomes a click, and the business gets the appearance of control without the substance.

This tendency has a name, automation bias: over-relying on what a system produces. It is serious enough that the EU’s AI Act names it. Article 14 requires that, for high-risk AI systems, the people assigned to oversee them are enabled to remain aware of the tendency to automatically rely or over-rely on the system’s output. The same article says oversight measures should be proportionate to the risks, level of autonomy and context of use. That law covers specific high-risk uses and is not a rulebook for everyday business automation, but it shows how seriously the problem is taken, and “proportionate” is the principle behind this guide.

Having the AI explain itself doesn’t fix the problem, and may worsen it. In a 2025 conference paper, researchers ran a field experiment in which 228 evaluators screened 48 early-stage innovations, with no AI, with AI recommendations alone, or with recommendations plus written explanations. Both AI conditions beat human-only screening. But the explanations made reviewers more likely to go along with the AI without making their decisions any better than the plain recommendation, and there were tentative signs that they increased rejection of promising unconventional ideas. It is one study on one kind of task, so treat it as a caution, not a law. The reference is in the Sources below.

The opposite approach has an obvious cost. If you approve nothing, the first expensive mistake happens without anyone knowing until a customer or your accountant tells you.

Four levels of freedom

Instead of a yes or no, think of four levels. You can give each action in a workflow its own.

1. Act alone. The AI does it and nobody is told. Suitable for low-stakes, easily reversible tasks such as filing an email or tagging a record.
2. Act and tell. The AI does it and a person is notified, or sees a daily summary. They can step in afterwards, but nothing waits for them.
3. Draft and ask. The AI prepares the action and stops. A person approves, edits or rejects it before anything happens.
4. Suggest only. The AI offers an option or an analysis, and a person does the work and makes the decision.

You may also hear the terms “human in the loop” (a person approves each step) and “human on the loop” (the AI acts and a person monitors). Level 3 corresponds to the first, and levels 1 and 2 to the second. The point of the four levels is that the choice is a scale, and each action belongs somewhere on it.

Four questions to place any action

For each action the AI might take, ask:

Can it be undone? A mislabelled file can be fixed in seconds. A sent message or a transferred payment cannot be taken back, or only at some cost.

How bad is it if it’s wrong? Think about money, customer trust, legal exposure and the effect on staff. A small error with a cheap fix is very different from a wrong figure in a client’s invoice.

Would you notice? Some mistakes are visible straight away. Others, like a slightly wrong record in a database, can sit unnoticed for months. The less visible the mistake, the more reason to look in advance.

How often does it happen? High volume makes human review expensive and unreliable. That is a reason to automate more of the safe actions and concentrate approval on the few that matter.

Irreversible, costly and invisible points toward level 3 or 4. Reversible, cheap and obvious points toward level 1 or 2. Here is how it works for some common actions (an illustrative guide, not a rule):

Make the approval worth something

If you decide a person must approve something, design the approval so it actually works.

Ask for fewer approvals. Concentrate human attention on the actions that really need it. Every unnecessary approval trains people to click without reading the important ones.

Show what changed and what’s unusual. An approval screen that says “Approve payment?” invites a reflex click. One that shows the amount, the supplier, how it differs from the last six payments and why it was flagged invites a decision. Keep the AI’s own reasoning secondary, given what the research above found about persuasive explanations.

Review meaningful units. A complete proposed reply, or a whole batch of related changes, gives the reviewer something to reason about. A stream of tiny fragments encourages pattern-matching.

Name the person who owns it. Whoever approves should know the decision is theirs, not a formality. When a person and an AI both touch a decision, it can become unclear who answers for the result, so say it out loud in the process.

Test whether anyone is reading. Occasionally slip a deliberately wrong item into the queue and see whether it is caught. If it isn’t, the control isn’t working, and you want to find out before a real mistake does it for you.

Don’t depend on the AI knowing the rules. If an agent must not send money, the safest way to enforce that is in the system’s permissions, not in an instruction telling it to ask first. An AI can misread or ignore instructions, but it cannot use an access it was never given.

Earn autonomy, don’t grant it on day one

You don’t have to settle these levels permanently. A sensible approach is to start strict and relax with evidence.

Begin with a person reviewing nearly everything. Record how often they change or reject what the AI produced. If an action is almost never corrected over a few weeks, move it up one level, from “draft and ask” to “act and tell”, for example. If corrections stay frequent, leave it where it is and look at why.

Set triggers for moving back too: a customer complaint, an unexpected error, a change in the tool or its inputs. Autonomy that was earned can be withdrawn, and a quarterly review of the levels keeps the setup honest.

Before you hand any process to AI at all, it’s worth checking whether the process itself is understood. Our article on why automation projects fail explains how hidden, undocumented work in a process causes automations to go wrong. And if you’re weighing how much independence to give an AI agent specifically, our post on AI agents vs traditional automation explains how they differ in what they can do on their own.

Frequently asked questions

Decide each action on its own

The right amount of oversight isn’t a company-wide setting. It differs for every action, and it should change as you learn how the automation behaves. If you’d like help deciding where those lines belong in your own processes, our Consultancy and Training service works with teams on exactly that, and you can get in touch to start the conversation.